Goal of this talk

• Convince you to invest in large scale lab based performance analysis.
• Dig into some data.

Overall Distribution

Overall Distribution (Sampled)

Objective of this analysis.

• Find Chrome and Website bugs based on new insight.
• Understand Page Load performance at scale.
• Identify weaknesses in our metrics.

Gathering Data

• Load the Alexa top 10k on Nexus 5’s.
• Simulate 3G network.
• Both warm & cold Page loads.
• Reasonably detailed main thread attribution.

Time To Consistently Interactive

Time To Consistently Interactive | Quantiles

Time To Consistently Interactive | Normalized by quantiles

Important Timestamp Deltas

Per Second Contributors

Next Steps: Elimate Other

• Issue.
• Some low hanging fruit:

Slice Title: Total self time in ms across top 10k pages
TaskQueueManager::RunTask : 45799357.7984
LocalFrameView::prePaint : 22831219.588
ProxyMain::BeginMainFrame::commit : 16157595.614
TaskQueueManager::ProcessTaskFromWorkQueue : 16034099.1022
V8.NewContext : 15784434.0443

Next Steps: Elimate Other

• We need some better ways to keep trace event renaming from breaking metrics.

• Contribute by adding your slices to the user friendly category driver.

  const TITLES_FOR_USER_FRIENDLY_CATEGORY = {
  composite: [
    'CompositingInputsUpdater::update',
    'ThreadProxy::SetNeedsUpdateLayers',
    'LayerTreeHost::UpdateLayers::CalcDrawProps',
    'UpdateLayerTree'
  ]
  }

Next Steps: Improved Breakdowns

• blocked_on_network should split out by type of resource.
• Can script_execute be split out via runtimeCallStats?
• Filter out 404s - maybe pick a better pageset?
• ???

Next Steps: Crazy Ideas

• Label sites by framework.
• Gather memory data as well.
• Do the same thing, but for input.